Insight is often romanticized as a “flash of genius,” but its true mechanism is much more structured and computational. It is neither such a flash nor, on the other hand, the result of mere rule-based symbolic manipulation of expressions. I will argue that it comes from a learned, constrained search through a space of representations which yields a reduction in complexity. Further, this search is what underlies intuition. Consider the telling example of a chess grandmaster choosing a move to make in a game. As world champion Magnus Carlsen attests, he does not mechanically consider the position’s possible moves and calculate many moves ahead. He is usually only conscious of a set of clearly good moves and chooses from among them (Shukla, 2026). This ability comes from years and years of statistical learning, as will be explained. My argument will first focus on the nature of intuition in practice, then develop into a discussion on representations, complexity, and mathematical insight, and finally will incorporate ideas and empirical findings in artificial intelligence and biology.
Throughout this essay I will employ the word “insight” in a similar way as one would employ the word “creativity”: the capability of producing novel, surprising, and valuable solutions to problems (Halina, 2021). This is in contrast to the other way it is used in sentences like “I read X and got insight into the affairs of Y.” I will also speak about representations, by which I mean internal encodings of an object that capture salient information but are less complicated than the object. For example, a representation of the irrational number 3.1415926… is simply the ratio of the circumference to the diameter of any circle. I will now turn to the main argument about intuition and insight.
Someone with strong intuitions in a field knows when to perform what actions in highly complicated scenarios. An intuitive soccer player knows where to lob the ball in a cross, and an intuitive artist knows exactly what basic colors to mix to get a desired hue. Formally, we can describe an agent’s intuition as a function that takes in a broad range of varying inputs and quickly and accurately outputs some action to take. Intuition is not the mapping to the output, but the traversal to that output. Consider again the chess player. Chess, famously, is an enormously complicated game and admits roughly 10^120 different possible games (according to Claude Shannon’s estimation, 1950), so it is essentially impossible to play skillfully by brute force considering all possible moves (then recursively considering all opponent responses, then all responses to that response, etc, until checkmate). Chess requires more efficient ways to capture the salient information about a position. A chess player’s intuition is related to the process of scanning the board state and representing a position in simpler terms (not the actual move that the player decides on): when a player considers representations such as “central knight” and “strong pawn structure” they are intuitively navigating the space of board representations which directly imply what move to make for the purpose of winning the game; that they are navigating a connected structure is evidenced from the fact that these representations heavily interact, for example: central knights are very strong when the opponent has a weak pawn structure.
A chess novice might only consider which of his pieces are under attack and move them out of the way accordingly. He is representing the board as a series of paths his opponent’s pieces can take. His intuitive process is a very simple calculation that checks these representations and accordingly takes action to move his own pieces out of the way. Of course, this representation scheme is much more efficient than recursively calculating moves and moves ahead, but it leaves much to be desired. More learned players think at a more abstract level. In addition to the concerns of the novice, they begin to take into consideration whether they can attack two pieces at once, whether their pawns are blocked from moving, and whether their knight is centralized. They also need to understand how these representations of a position interact with each other: “ceteris paribus, move the knight to a central square if it will not block other pieces.” The most intuitive chess grandmasters in the world will, by their own account, operate through even more abstract representations, such as “Passive Knight-Dominated Middle-Game with a King's Indian Opening Structure.” They have learned, through thousands and thousands of games, to distill so much information about what kinds of considerations and moves lead to their winning a game that such complex representations pop out to them in the same way that an attacked queen pops out to the novice.
When guided by intuition, one may find an insightful solution; in the above example, this might be a brilliant queen sacrifice. This occurs when one finds a representation that massively reduces complexity and reveals previously hidden goal-directed actions. Like representations of chess positions, mathematical statements admit different equivalent representations that describe the original statement from a new perspective and hence admit the same truth values. Consider the simple mathematical puzzle: “How many ways are there to select n − 1 people from a group of n different people?” Put as a statement: “Prove that there are n ways to choose n − 1 subjects from a group of size n.” One proof might be to list out every combination of subjects and sum them up. A more insightful proof might be to realize that picking n − 1 subjects is equivalent to choosing 1 subject not to pick, of which there are clearly n possibilities. This is insightful because it is novel and surprising in that it does away with the fixed tendencies to view mathematical objects as serving particular purposes and imagines them used in alternative ways.
This solution seems to have qualitatively reduced the problem’s complexity. There is a quantitative way to show this. Kolmogorov defined a way to measure the complexity of an object X as the length of the theoretically shortest possible description that produces X in a given language (be it binary code or modern English). That measure belongs to the object, but a similar measure for the proof is conditional on it: how long a description the solver needs, given the representation it currently holds. Written out in full, the n groups of n-1 people take on the order of n^2 names; the new perspective of whom to leave out can be described by a list of n values. It is this reduction in description length that allows the solution to pop out. The insight or “aha moment” is simply a transition from a poor representation to a rich one, and the recognition that it massively shortens the description and has promise toward the goal of, in this case, proving a theorem.
In mathematical practice, genuine insight does not come from symbolic manipulation of problems but rather from re-representing the problem in a compressing framework in which the solution, or path to the solution, becomes clear. For many mathematicians, a necessary feature of math proofs is that they reveal why a statement is true (Horsten, 2007). Sometimes, the crucial representational shift is not just a rearrangement of current symbols but a leap to an entirely different conceptual framework. As the algebraist Paolo Aluffi remarks on proofs:
“The reason we don't need to do any work now to recover the same fancy statement is that the steps in that proof have been absorbed by elements of the language we have learned.... Rather than devising clever proofs, one tries to devise a clever language, and then coming up with proofs is more or less the same as coming up with (grammatically correct) sentences in that language, something our brains are evolutionarily adept at doing" (Aluffi 2021).
In the terms used above, a clever mathematical language lowers the conditional complexity of every proof written in it. Many historical mathematical breakthroughs only occurred by developing an entirely new framework for mathematical objects. Extending “real” numbers (those you might find on an infinite number line) to the complex plane (which involves perplexing numbers like the square root of −1) has allowed for fundamental breakthroughs from techniques to solve previously unsolvable integrals to major advancements in quantum physics.
However, it should be noted that not every insightful representation is necessarily an immediate compression. For instance, we sometimes think about two-dimensional mathematical problems in a projected three-dimensional environment, which is clearly the opposite of a compression. But from three dimensions it is sometimes clearer how to proceed with the proof, which will ultimately lead to more and more compressions, hopefully to the goal of entirely reducing complexity to “true” or “false.” The important feat is that we are reducing global complexity, just at the cost of a small local increase.
This account of intuition and insight is reflected in the working of modern machine-learning programs. Google’s artificial intelligence model AlphaGo shocked players and researchers by learning to play the notoriously complicated game Go after statistically synthesizing millions of games, and beat the world champion Lee Sedol. In fact, the Korean Go Association considered it so creative and insightful that it awarded it the highest rank of “9 dan.” Faced with a game position, AlphaGo relies on a statistically learned neural network “policy function” and “value function” to choose a move as follows: the policy network narrows the breadth of the search of moves by choosing high-probability moves, while the value network reduces the depth of the search by providing information about the value of a position (Halina, 2021) (Silver et al., 2016). This very closely matches the human case. The learned policy network corresponds to intuition: the learned pathfinding through a representational space of board positions, and the value network that determines how promising a position is for the objective of winning the game corresponds to further compressions of that representation to a numerical value.
The newest paradigm of human designed artificial intelligence is Large Language Models (LLMs). An LLM is an artificial neural network that is “trained” on a massive corpus of human text with an architecture that allows it to learn to produce linguistically meaningful and responsive sentences. It represents words or ideas in a very high-dimensional space that is designed to encode similarities to other words or ideas via simple mathematical operations. For years, LLMs on their own tended to produce plausible seeming but incoherent arguments, replete with hallucinatory references to theorems and circular arguments. State-of-the-art math-reasoning models that used LLMs augmented them with formal proof assistants, mechanical self-verification algorithms, and lots of post-processing steps (Asperti et al., 2026). That changed at a breakneck pace. In 2024 LLMs could solve SAT math problems; in 2025 a version of Gemini produced rigorous natural language proofs on the International Mathematical Olympiad (Luong and Lockhart); in 2026 OpenAI announced an AI proof regarding the million dollar Navier-Stokes problem that had been unsolved for decades (OpenAI). As the account of insight would suggest: informal proofs and reasoning align much better with current LLMs than formal proofs and reasoning. In the former case, reasoning patterns appear in text the LLM is trained on and are internalized through statistical learning. This is because coherent narratives, natural language, conventions, and approximations are precisely the kinds of text that LLMs are trained on, and training modern transformer-based models tends to result in the kind of high-dimensional conceptual representation space that is needed. In the latter case, however, every iteration of logic must follow exact brittle rules of inference. No gaps can be left, as tiny errors can render formal proofs completely invalid. This is difficult for LLMs to obey without explicit feedback, unlike in informal math in which association, rather than deduction, is the most important. Interestingly, original AI systems, sometimes called Good Old-Fashioned AI, were designed with explicit, formal symbolic-manipulation rules based on Alan Turing’s ideas about Turing Machines. They excelled at formal deduction but were very poor at the very things LLMs are good at: learning complicated associations between words and ideas that allow them to use analogies and navigate high-dimensional representations.
What remains difficult is the deeper kind of insight: inventing a new representation and formalism rather than navigating learned ones. Zeng, Griffiths, and Lake (2026) asked whether small language models could discover the concept of zero from arithmetic that never included it. They could not discover the concept from scratch at test time. Whether this is a structural problem with LLMs or merely the next milestone they will reach has yet to be determined.
Becker et al. (2025) sought to investigate the neural correlates of insight, specifically the sort of “aha” moment reflected when one suddenly recognizes the contents of a noisy, blurry image. Here, insight would be the compression from random noise (whose complexity would be the description of the contents of every pixel) to, for instance, the blurry face of a celebrity (which has a very short English descriptor). It is thought that representational change in the brain involves a change in the pattern of neural connections in an associated region. The researchers focused on the ventral occipito-temporal cortex of the brain, which is associated with complex visual object recognition. They compared activity patterns before and after the image was suddenly recognized. Strikingly, they found significant evidence that the sudden recognition was correlated with statistical measures of reorganization of cortical representations. This lends strong evidence to the psychological description of insight being a kind of representational shift (Becker et al., 2025).
One might still think that insight is more like a sudden flash than some navigation through a representation space; indeed, it does feel subjectively as though insightful thoughts simply pop out of nowhere. But I would argue that introspection of this sort is often misleading. To illustrate this point, consider the nature of recognizing, say, a cup. Most would agree that it feels as though there is some atomic “cup module” that activates in the brain when a cup is in the visual field, but empirical findings have shown that vision operates in a highly complex hierarchical network in which first edges are recognized in the visual cortex, then local features like a handle, then the prefrontal cortex repeatedly provides predictions about the object and refines them until it has a strong model of it, and finally the semantics of “cup” are passed to the anterior temporal lobe. Researchers failed to understand neurobiological processes like vision from pure introspection, so it should not a priori be the case that the natures of intuition and insight can be introspected accurately.
To summarize, intuition, whether in mathematics, chess, or neural networks, is fast navigation in a learned representation space. Insight occurs when intuition points toward a new representation where the problem compresses, revealing a solution that was opaque before. Considering that intuition and insight are crucial in advanced reasoning abilities, focusing our efforts on encoding them in AI systems could translate to more and more intelligent systems. A major goal of AI research is to develop “the kind of AI we need for science,” as Demis Hassabis, the head of the group that produced AlphaGo, remarked. Another promising direction research might lead to is strong hypothesis-generation in the more empirical sciences. Lu et al. (2024) demonstrate a model that, given a starting template, “brainstorms” a slew of novel (and perhaps insightful) research directions that are then taken up by other artificial systems to produce research and research papers that established scientists tend to judge to be somewhat important and novel (Lu et al., 2024). The grandest hope is that the more that we can incorporate insightful artificial intelligence into the enterprise of scientific research, the more we can advance medicine, answer ancient questions, and improve the human condition.
Citations
Aluffi, Paolo. Algebra: Notes from the Underground. Cambridge UP, 2021.
Asperti, Andrea, et al. “Thinking Machines: Mathematical Reasoning in the Age of LLMs.” Big Data and Cognitive Computing, vol. 10, no. 1, Jan. 2026, p. 38. Crossref, https://doi.org/10.3390/bdcc10010038.
Becker, M., Sommer, T., & Cabeza, R. (2025). Insight predicts subsequent memory via cortical representational change and hippocampal activity. Nature Communications, 16, Article 4341. https://doi.org/10.1038/s41467-025-59355-4
Halina, M. (2021). Insightful Artificial Intelligence. Mind & Language. https://onlinelibrary.wiley.com/doi/10.1111/mila.12321
Horsten, Leon, “Philosophy of Mathematics”, The Stanford Encyclopedia of Philosophy (Winter 2023 Edition), Edward N. Zalta & Uri Nodelman (eds.), URL = https://plato.stanford.edu/archives/win2023/entries/philosophy-mathematics/
Lu, C. et al. (2024). The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery. arXiv:2408.06292. https://arxiv.org/abs/2408.06292
Luong, Thang, and Edward Lockhart. “Advanced Version of Gemini with Deep Think Officially Achieves Gold-Medal Standard at the International Mathematical Olympiad.” Google DeepMind, 21 July 2025, deepmind.google/discover/blog/advanced-version-of-gemini-with-deep-think-officially-achieves-gold-medal-standard-at-the-international-mathematical-olympiad/.
OpenAI. “On the Navier–Stokes Millennium Prize Problem.” OpenAI, 8 Sept. 2026, openai.com/index/navier-stokes-solution/.
Shannon, C. E. (1950), Programming a computer for playing chess, Philosophical Magazine, 41(314).
Shukla, A. (2026, March 3). Magnus Carlsen reveals how many moves ahead he calculates, busts biggest chess myth. The Sunday Guardian. https://sundayguardianlive.com/sports/chess-news-magnus-carlsen-reveals-how-many-moves-ahead-he-calculates-busts-biggest-chess-myths-chess-grand-master-173703/
Silver, D., Huang, A., Maddison, C. et al. Mastering the game of Go with deep neural networks and tree search. Nature 529, 484–489 (2016). https://doi.org/10.1038/nature16961
Zeng, Phoebe, et al. “Nothing from Something: Can a Language Model Discover 0?” arXiv, 15 June 2026, arXiv:2606.17289.